Rethinking Explanations: Formalizing Contrast in Description Logics

Explanations need to match how people actually think: The paper argues that current AI explanation methods only show reasoning steps, but users often want to understand why one outcome occurred instead of another. For example, a person might ask "why was my loan application denied instead of approved?" rather than just wanting to know the steps that led to denial. This user-centered approach is more aligned with how people naturally seek explanations.

Context and prior expectations matter for understanding: The research is motivated by the observation that users are often surprised by outcomes because they anticipated something different. Effective explanations should address this gap between what happened and what was expected, making AI systems more intuitive and trustworthy to non-expert users.

Contrastive explanations reveal meaningful differences: Rather than providing separate explanations for "why P is true" and "why Q is false," the paper proposes showing the actual differences between two scenarios. This approach helps users understand not just individual facts, but the critical distinctions that led to different outcomes—potentially making AI decision-making more transparent and easier to scrutinize.

This framework applies to real-world decision systems: The formal foundations developed for description logics can be implemented in knowledge-based systems that make decisions (like loan approvals, hiring, or eligibility determinations), suggesting practical applications for making AI systems more explainable in high-stakes domains.